Kernel-grade FinOps.
For AI fleets.
A scoped memory-wall audit, typically 1–3 weeks. We hook sys_exit_ioctl and measure the nanosecond gap between every MEMCPY at the driver level with zero observer-effect latency. The output: a savings PDF showing exactly how much of your fleet is memory-bound, not compute-bound, what to quantize, and what to reclaim.
- ✓On-prem · read-only · no agent SDK installed
- ✓Roofline classification · per-PID, per-GPU, per-shard
- ✓Sharding-aware · auto-detects Tensor Parallelism (TP) degree
- ✓vLLM-aware · KV-cache vs weight-load distinction
- ✓Quantization advisor · FP16 → INT4 / GGUF guidance
- ✓Output · savings PDF · roofline + waste score + reclaim plan
See what your fleet is actually losing.
The Efficiency Audit is the low-bar way to start with Arca. It works for any company running a real GPU fleet on hardware it controls — owned or dedicated / bare-metal rented. In read-only mode we measure how much of your spend is going to idle allocation, orphaned contexts, and memory-bound stalls, and hand you a signed report of what you can reclaim.
The deliverable is a signed savings report: a roofline classification of every workload (compute-bound vs memory-bound), a quantization advisory, and a per-shard VRAM reclaim plan — measured against your traffic, not a template.
For example, a fleet whose dashboards report 71% utilization may be doing real compute for only a fraction of that time, with the rest spent waiting on memory. The audit quantifies exactly how much, and what it is worth to reclaim. Every figure in your report is measured on your hosts — illustrative numbers shown elsewhere on this site are not results.
The same engine runs continuously inside Arca Sentry: the audit is Sentry in read-only mode, so what you measure is what you can later choose to enforce.